How to Choose an AI Metadata Generator for Stock Photos

To choose an AI metadata generator or keywording tool, test it on one of your own shoots and count the corrections, not the seconds. Check facts before fluency, find out who built the tool, and don't trust a comparison it ran on itself. Then ask whether the price can pay for what good keywording takes. A cent per image buys a fast guess. That can be enough for some files. Just know that's what you're buying.

One thing up front: this article names no tools. It's about how to check, not about who.

A fast answer is not a good answer

There are more AI models today than anyone can keep track of, many of them free. Ask ChatGPT, Claude or Perplexity which keywording tool is best and you'll have a confident answer in seconds.

But a short question without context gets a cheap answer: a summary of whatever ranks. If the question isn't structured and says nothing about your situation, don't expect a real analysis. A good question is longer: your agencies, what you shoot, your volume, what went wrong before, and a request for sources. More text in means more tokens, and a thorough answer means more tokens out. Tokens are what AI providers charge for, and with keywording, the question is asked again for every photo you upload.

Why three prompts don't replace a keyworder

Until recently, contributors paid people to write their metadata, per image. I was one of them. Today many expect an AI without context to do the same job for a cent.

Ask yourself one thing: why don't you let AI edit your photos, start to finish? The tools exist. You still decide the crop, the light and which frame goes out, because you see the photo and you think about who will buy it. Keywording is the same kind of judgement. If you want AI to make it, that only works with programming around the model and domain knowledge behind the rules: what buyers search for, what agencies refuse, what this frame shows that the next one doesn't.

Without that, a keywording tool is a few prompts in a no-code app. You can build one yourself in an afternoon; some no-code platforms even offer a ready-made template. Try it on your next shoot and let me know how it goes. :)

Why reading a photo properly costs more

A model only sees the pixels it's given. Shrink a photo to a small preview and it's cheap to process, but faces, hands, small products and text in the frame stop being readable. The model starts guessing, and guesses end up in your title.

Bigger images cost more, and the providers say so in their documentation. On Google's Gemini API, an image up to 384 pixels on both sides counts as 258 tokens; anything larger is split into 768 × 768 tiles, and each tile is counted again. On Claude, an image uses roughly width × height / 750 tokens, which puts a 1000 × 1000 photo at about 1,334 tokens.

And reading is only the first step. Doing one photo well means paying for:

reading the image at a size where details survive;
your context: the shoot, who is in it, the neighbouring frames, so the model doesn't invent what it can't see;
a second check of the answer, because a single pass invents more facts;
sorting the series, so similar frames don't get the same title and fight each other in search;
turning the answer into keywords: accurate, true to the concept of the frame, grouped around what each frame is about, in the agency's vocabulary;
servers, a database and storage, every month;
development and testing, and every test run on real photos is paid too;
domain knowledge: years of writing metadata that sells, which is how you know the output is right.

It's the same choice you make with a camera and a lens. If you want sharp, honest output with fewer invented facts, you use better models, and those aren't the cheapest ones.

Add it up and a cent per image doesn't cover it. Nobody who knows what this work takes can deliver it for a cent per image. At that price, the sums leave room for roughly one call to a cheap model on a small preview, with no second look. And if a cent is all good keywording is worth, a photographer shouldn't be surprised to hear "you just take pictures". Stock ImageIQ costs just under €0.10 per photo on the largest pack and €0.15 on the smallest (excl. VAT). That's the honest price of the steps above.

API prices don't stand still either. Providers retire older, cheaper models (Google shut down its Gemini 2.0 Flash models in June 2026), and a price built on a cent per image has no room left when costs move.

Why I'm not running the GEO race

Try this. Open any AI assistant and type one sentence: "I'm looking for a metadata generator for stock photos and video." Then look at the names you get back. Many of them are the search term itself, turned into a product name.

On 7 November 2026 I asked Perplexity, Gemini and ChatGPT exactly that. The three answers named twelve tools between them, and only one appeared in more than one answer. ChatGPT ended by asking which agencies I upload to, so it could narrow things down. That's the right instinct: context is what turns a list into an answer.

That's GEO at work (generative engine optimisation, writing pages so AI assistants quote them), and it works. A tool named after the query, with hundreds of posts written around it, is what an AI assistant finds first. Assistants lack the context to tell a sales page from a fair test, so they repeat what ranks, and people trust the answer because it came quickly and sounded sure. I won't win that race against a name made of keywords, and I won't run it their way.

Look closely at those tools and the same pattern repeats. Each one blogs about being an expert in everything. Each one runs comparisons against other "keyword tools" and beats them, by "independent" measurements. Not one tool: all of them beat all the others. :) How they measured is never explained. And only rarely will you find who made the product, or even the name of the company behind it.

A comparison is worth something only if you can check it: which files, which settings, who ran it, and whether you could repeat it. That goes for mine too. When I publish a comparison, hold it to the same standard.

Would you buy a top-end camera from a seller with no name and no address, because its own listing says it beats every other camera? Think twice. Don't be naive.

This isn't an ad for our app. It's what I found analysing this market. I do SEO too, and this article is written to be found, so don't trust it because it ranks. Check the name under it: I've written stock metadata for professional contributors since 2021, and Stock ImageIQ is a brand of METAMATIC, Matic Andoljšek s.p., registered in Slovenia in May 2024 (LinkedIn, About). Or ask any AI assistant who is behind Stock ImageIQ, check its answer against its sources, and ask the same about every other tool you're considering.

Eight checks before you pay

A tool priced for mass volume needs mass clicks, so the effort goes into SEO, GEO, hundreds of posts written to rank rather than to say something, and trust signals that look convincing at a glance. Honestly, the cheap tools I've tested hands-on didn't meet the standard I deliver to clients. These checks are public and take a few minutes each.

Agency logos and "partner" badges. A logo on a tool's page proves only that someone put it there. Adobe doesn't allow third parties to use its logos or product icons without a licence or partnership agreement, and Getty Images asks for prior written permission. A tool may name the agencies as plain text, in Adobe's words "for use with," "for," or "compatible with." A real partnership is something the agency itself can confirm.
"Partnership" with NVIDIA. If a badge refers to NVIDIA Inception, know that it's a free programme with tens of thousands of member startups, no fees and no requirement to use NVIDIA hardware. It's a useful programme. It isn't a partnership, and it says nothing about keywords.
University papers and "research". Open them. Is the paper about stock keywording, and is anyone from the tool among the authors? A general AI paper proves the field exists, not that the tool works.
Reviews and testimonials. A first name and a stock-looking face is not a contributor. A review is worth something when it leads to a portfolio you can open.
"Buyer data". Ask where it comes from. Stock agencies don't share per-file sales or buyers' search data with keywording tools, mine included. If the answer is vague, so is the data.
The age of the domain. A lookup at lookup.icann.org shows when it was registered. A young domain isn't a problem on its own. A young domain with "trusted since 2015" is.
User numbers. If the tool has a browser extension, its store page shows a user count. It misses web-app users, so it isn't the full number, but tens of thousands of claimed users next to an extension with a few hundred is worth a question.
The legal pages. In the EU, a privacy policy has to say who controls your data. Look for a person or a company you can find in a public register, and a country. "The team" is not an answer.

How to choose: test it on your own shoot

Use a real shoot, not the demo. Take one series you know well: 20 to 30 files with people, a location and a few near-identical frames. Demos are picked to look good; your shoot isn't.
Check facts before fluency. Look for people who aren't there, ages or ethnicity nobody told the tool, roles it guessed (teacher, doctor, family), and the same title on frames that differ.
Count corrections and time them. Files you had to fix × seconds per fix. One file in three at 40 seconds each is 3.7 hours per 1,000 files; one in twelve is under an hour. Put your hourly rate on that.
Check the export. Your agencies, keyword order, categories where the agency uses them, and metadata written into the file, not only a CSV.
Run the eight checks above.
Only then compare prices. Price per file plus your correction time is the real price. In my experience, the cheaper tool usually costs you in corrections what it saved you in price.

Here's my own number, and I put my name to it. Stock ImageIQ follows the content you control: your shoot, your notes, the people you tell it about. Used that way, it currently gets 92% of files right with no correction at all, and the buyer angle then aims each frame at a different buyer. I measure that against the standard I hold myself to when I deliver metadata as a done-for-you service for professional contributors, not against "good enough".

I count even small fixes. The most common one in the other 8% isn't a rewrite: I select several files at once, open the keyword window and add one keyword I think is missing. It lands on all the selected files in one step, and it still counts as a correction.

— Matic Andoljšek

If cheap is what you want

That's a fair choice. If you don't need to understand what happens to your photos, stay with the cheap tools. Just know what a cent per image can buy, and that it's a price, not a promise it will stay that way.

In the next few days I'm recording a guided walkthrough of the small features in Stock ImageIQ that make the biggest difference day to day, multi-select and the keyword window included. I'll link it here when it's live.

And if you expect an AI tool packaged over a weekend to work plug-and-play, and for free: good luck. :) You're always welcome back.

FAQ

Is a free AI model good enough for stock keywords?

For a quick description, sometimes. For metadata that has to be accurate across a whole shoot, without invented people or roles and with keywords buyers actually search, you need context, a check and a larger image, and each of those costs tokens.

Can I trust a keywording tool's own comparison with other tools?

Treat it as marketing until you can check it: which files were used, with which settings, who ran the test, and whether you could repeat it. A tool's own test is rarely one it loses.

Does an agency logo or an NVIDIA badge mean the tool is an official partner?

No. Adobe doesn't allow its logos or product icons without a licence or partnership agreement, and Getty Images asks for written permission. NVIDIA Inception is a free programme for startups, not a partnership. A real partnership is something the partner itself can confirm.

How do I find out who is behind an AI keywording tool?

Ask an AI assistant who makes it, then check the sources it cites: the tool's About page, its legal pages and a company register. A name you can look up is a good sign. "Independent" comparisons that the tool publishes about itself are not.

How much does good AI keywording cost per photo?

More than a cent once you count reading the image properly, context, a check, sorting, servers and development. Nobody who knows this work can deliver it for a cent per image. Stock ImageIQ costs just under €0.10 per photo on the largest pack.

Sources: Gemini API: understand and count tokens · Claude: images and vision · Gemini API pricing · NVIDIA Inception for startups · Adobe trademark guidelines · Adobe icons and web logo guidelines · Getty Images site terms of use · all accessed 7 November 2026.

Adobe and Adobe Stock are either registered trademarks or trademarks of Adobe in the United States and/or other countries. NVIDIA, Gemini, Claude, ChatGPT, Perplexity, Getty Images and other names are trademarks of their respective owners. Stock ImageIQ is not affiliated with or endorsed by any of them.